| An adaptive power system stabilizer based on recurrent neural networks is developed in this dissertation. The Real-Time Recurrent Neural Networks with the Real Time Recurrent Learning (RTRL) algorithm are applied in the design of a recurrent neural network based power system stabilizer (RNN PSS). The structure and training procedure of the proposed RNN PSS are discussed.;The architecture of the proposed RNN PSS has two recurrent neural networks. The first one functions as an identifier to learn the dynamic characteristics of power plant, the second one functions as controller to damp the oscillations of power plant caused by different disturbances. The training of these two neural networks has two stages: off-line training and on-line update. There is no reference model needed for the proposed RNN PSS. It is trained directly based on the input and output of the plant.;Simulation studies and comparison between the proposed RNN PSS and the conventional PSS are conducted on both a single-machine infinite-bus power system model and a multi-machine power system model. The results demonstrate the effectiveness of the proposed RNN PSS in damping oscillations in the power system.;The designed RNN PSS is also implemented on a TMS320C30 Digital Signal Processing Board in a real time environment, and applied to a physical power system which consists of micro-alternator. DC motor, ABB PHSC2 PLC. The laboratory test results show that the prototype RNN PSS can provide a good response compared to the conventional PSS. |
